NVIDIANVIDIA-Certified Professional: Agentic AI2026-102.6 Evaluate and refine agent decision-making strategies.
Choose description, prompt, model, or RL fixes
Completed by the knowledge check
Learning objective
Refine action selection via tool descriptions, prompts, model choice, and RL fine-tuning.
Core idea
Refining action selection means changing the component responsible for a bad decision. If two tool descriptions overlap, clarify their distinction. If the agent repeatedly ignores a prerequisite, refine its decision instructions. If it cannot follow the interface reliably, compare compatible models. If a stable, verifiable task distribution still defeats those changes, consider reinforcement-learning fine-tuning.
These interventions change different artifacts and carry different costs. Use the measurements from 2.6.1 to locate the failure, and the controlled comparisons from 2.6.2 to verify the repair. See 2.1.1 and 2.3.1 for prompt and tool construction; this lesson decides which artifact to change after observing wrong actions.
How it works
Repair ambiguity at the tool interface
In NeMo Agent Toolkit, FunctionInfo.from_fn(..., description=...) supplies a function description; the documented alternative derives it from the callable’s docstring. Change this exposed description when the agent confuses neighboring capabilities. State when to use the tool, its distinguishing inputs, and the result it returns. Describe a status lookup as read-only and distinguish it from a tool that changes status.
Test the pair of tools on minimally different requests: “show the schedule” and “change the schedule” should not select the same action merely because both mention a schedule. Include missing-input cases and plausible synonyms. Measure the confusion between specific tool pairs, rather than accepting a higher overall answer score as proof that the ambiguity disappeared.
For the ReAct agent, include_tool_input_schema_in_tool_description defaults to True and appends input schemas to descriptions. If an argument is missing or unclear, check whether the flag was disabled or the schema field is poorly described. Adding examples to the system prompt before checking the exposed description risks duplicating or contradicting the tool interface.
Refine the decision instruction
Use the ReAct agent’s additional_instructions for decision guidance that belongs beside its base instructions. For example, require a lookup before deciding whether the requested operation is feasible, and require clarification when the entity identifier is ambiguous.
If you override system_prompt, preserve the tool variables and required ReAct output format described in the agent documentation. A prompt that improves business reasoning but breaks Action and Action Input parsing replaces one failure with another. Track parsing failures separately from incorrect choices after a valid parse.
Prompts guide model behavior; they do not confer permission to execute. Continue enforcing mutation authority in the adapter described in 2.3.2. A description saying “read-only” is useful selection information, while the tool implementation determines what the call can actually do.
Compare a compatible model
The agent’s llm_name references an entry under llms. Change that reference for a model comparison while keeping tools, prompts, fixtures, and result checks fixed. Measure whether the candidate improves the specific failing decisions and still meets the application latency and token budget.
A model substitution must support the selected agent’s interface. Native tool calling requires tool/function-calling support; a stronger text-generation benchmark is not a substitute for that capability. For a ReAct implementation, assess adherence to the required textual action format. Measure valid argument generation as well as tool identity.
Choose a model change when clearer information and instructions still leave capability failures. Do not pay its inference cost merely to compensate for two indistinguishable descriptions.
Automate prompt and model comparisons with nat optimize after installing the optional nvidia-nat[config-optimizer]; see 3.4.3 for the optimization mechanism.
Train against verifiable outcomes
NVIDIA documents a NeMo Gym and NeMo RL tutorial that trains multi-step tool calling with Group Relative Policy Optimization (GRPO) in a Workplace Assistant environment. This is a model-training path, separate from editing a workflow prompt. The training configuration’s num_generations_per_prompt controls the rollouts generated for each prompt; its illustrated value is four. Those extra rollouts consume inference work before the model update, so training cost includes environment executions as well as GPU training.
The tutorial connects Gym through env.should_use_nemo_gym: true and env.nemo_gym.config_paths. Separate data.train.data_path from data.validation.data_path to test action selection beyond the training examples.
Implement the verifier in a NeMo Gym resources server. With SimpleResourcesServer, the agent can initialize per-task state through /seed_session and score the completed rollout through /verify. A reward for syntactically valid JSON can improve formatting while leaving the wrong action untouched. A reward for task completion alone can favor shortcuts unless the verifier also checks required prerequisites and forbidden state changes. The environment owner must make these conditions observable in the test state, not infer them from the model’s explanation.
The tutorial states that NeMo Gym itself does not require GPUs; GRPO training with NeMo RL does. For its Nemotron Nano 9B v2 run, the tutorial requires at least one node with eight NVIDIA GPUs of 80 GB or more each. Single-node testing uses one eight-GPU node; multi-node production recommends at least eight such nodes. This is the tutorial’s hardware scope, not a universal GRPO requirement. Adopt fine-tuning when you own representative tasks, a credible verifier, and the training operation. If the reward improves while held-out action correctness falls, inspect the verifier and task distribution before spending more training steps.
Key distinctions and trade-offs
Choose the smallest intervention that addresses the diagnosed cause:
| Intervention | Deciding condition | Mechanism | Cost and risk | Owner |
|---|---|---|---|---|
| Tool description | Confusion between similar capabilities | FunctionInfo.from_fn, description | More prompt content; descriptions can drift | Tool owner |
| Prompt | Missing decision rule or contrasting example | additional_instructions, system_prompt | Token cost; formatting regressions | Workflow owner |
| Model choice | Capability or interface failures persist | llm_name, llms | Serving cost and latency; model regressions | Model-serving owner |
| RL fine-tuning | Stable failures with verifiable outcomes | NeMo Gym environment, NeMo RL GRPO configuration | Rollout and training cost; reward shortcuts | Training and environment owners |
These are not a mandatory ladder. A missing permission check needs an adapter repair immediately. A workload with no reliable verifier is a poor candidate for reward-based training even if the team has spare GPUs.
Architecture / decision scenario
A media company’s publishing assistant chooses between previewing, scheduling, and publishing content. It handles 500 held-out editorial requests with a 6-second p95 target. An illustrative baseline produces 60 wrong tool choices: 35 confuse preview with publish, 15 skip a required schedule lookup, and ten persist after those defects are isolated.
The workflow owner compares all four interventions. A description revision reduces the preview/publish confusion to five cases. An additional_instructions revision reduces skipped lookups to three. A compatible model substitution reduces the remaining capability failures but increases p95 from 5 to 9 seconds, so the serving owner rejects it for the interactive workflow.
The tempting alternative is to train immediately on every recorded failure. The training owner instead requires a publishing resources server that initializes task state through /seed_session and checks the requested end state and unintended publication through /verify. The team reserves RL for a separate candidate and evaluates it on the untouched 500 requests with the same deadline and action checks. A higher training reward alone does not justify deployment.
The immediate choice is the description-and-prompt candidate, subject to the product’s acceptance criteria, with publication authority still enforced by the adapter. The unresolved failures remain visible; neither a slower model nor unverified fine-tuning is presented as an automatic cure.
Common exam traps
- Applying fine-tuning before identifying ambiguity in the exposed tools.
- Replacing a ReAct prompt without preserving its output interface.
- Assuming a more capable text model supports native tool calling.
- Rewarding valid syntax and calling it correct action selection.
- Treating a rising training reward as held-out task improvement.
Where the exam guide differs
The study guide’s suggested readings call the workflow toolkit “NVIDIA Agent Intelligence Toolkit (1.1.0)”; its documented name is NeMo Agent Toolkit.
What to remember
Use error categories to choose an intervention. Clarify tool semantics, repair decision instructions, compare compatible models, and train only against a trustworthy outcome signal. Verify every candidate on held-out action correctness and the serving budget, with operational permissions enforced outside the model.
Knowledge check
A ReAct publishing agent picks publish_article for requests asking to see how an article will look, though preview_article exists. Both tools’ descriptions say only “Handles an article.” Prompts, model and tools are otherwise unchanged. Which intervention addresses the cause most directly?
llm_name at a larger model with stronger benchmarksA ReAct agent’s tool descriptions are distinct and its arguments are valid, but it schedules posts without first running the calendar lookup the business requires. The base prompt has no such rule. Which change fits?
additional_instructionstool_namesllm_nameinclude_tool_input_schema_in_tool_descriptionadditional_instructions adds beside the base instructions. Removing the tool loses the capability, a model change pays for a failure clearer instructions may fix, and the schema flag affects argument information, not ordering.A team overrides a ReAct agent’s system_prompt to add business rules. Wrong-tool choices fall, but many runs now fail with parsing errors before any tool executes. What should the team do?
parse_agent_response_max_retries until the errors stopinclude_tool_input_schema_in_tool_description to shorten the promptA team moves a workflow to tool_calling_agent and shortlists the model with the highest score on a text reasoning benchmark. Which property must a candidate model have before its latency and quality are compared?
Action text formatreact_agent , and latency is compared only among models that can run the agent at all.A team trains a scheduling agent with GRPO through NeMo Gym and NeMo RL, rewarding each rollout that ends in a syntactically valid tool call. Training reward climbs, while correctness on held-out scheduling requests falls. What should the environment owner change?
num_generations_per_prompt to produce more rollouts per prompt/verify check end state, prerequisites and side effectsReferences
- NVIDIA, Writing Custom Functions
- NVIDIA, Configure the ReAct Agent
- NVIDIA, Configure the Tool Calling Agent
- NVIDIA, NeMo RL
- NVIDIA, NeMo RL Configuration
- NVIDIA, Gym Configuration
- NVIDIA, Build Verifiers
- NVIDIA, NVIDIA NeMo Agent Toolkit Optimizer
- NVIDIA, NVIDIA-Certified Professional: Agentic AI Exam Study Guide